摘要
[研究目的]针对现实社交网络中广泛存在的不实评论对谣言检测的负面影响问题,提出对抗学习框架下的谣言检测方法,从而在提升谣言检测准确率的同时,增强模型对噪声信息的容抗性。[研究方法]以信息对抗机制为基础,搭建具有融合结构及时序特征表示的生成网络,利用部分网络结构的共享及加强具有自注意力机制的二次鉴别网络,实现将非监督的对抗生成网络向有监督学习任务上的成功拓展。[研究结论]在PHEMEv5和新浪微博两个数据集上,该研究提出的模型在谣言检测的准确率上,相较于9种较为先进的基准模型至少提升了3.1%和4.1%;同时,实验显示,该研究提出的模型对于噪声信息并不敏感。充分证明了该模型在跨平台不同语言环境数据集上较高的谣言检测效果及较强的噪声容抗性。
[Research purpose]In order to cope with the negative impact of untrue comments on rumor detection,this paper proposes a rumor detection method under the adversarial learning framework,which can improve the accuracy of rumor detection and enhance the tolerance of the model to noise information.[Research method]Based on information campaign,a generative network is built with fusion of the structural and temporal characteristic representation networks.By means of the sharing of partial network structures,and the integration of the self-attention mechanism and secondary discriminative network,the unsupervised generative adversarial network is successfully extended to supervised learning tasks of rumor detection.[Research conclusion]On the two public datasets of PHEMEv5 and Weibo,compared with the nine typical baseline models,the accuracy of the proposed model is improved at 3.1%and 4.1%at least.Meanwhile,further experiments show that the proposed model is not sensitive to the noise messages on the task of rumor detection.The model proposed in this study improves the performance of rumor detection on cross platform datasets in different language environments,and shows a high tolerance for noise information.
作者
朱贺
Zhu He(Information Center of Library and Archives,Henan Normal University,Xinxiang 453007)
出处
《情报杂志》
CSSCI
北大核心
2024年第2期118-127,共10页
Journal of Intelligence
基金
2021年度河南省哲学社会科学规划年度项目“线上社交网络上舆情监测及引导策略研究”(编号:2021CZH021)
2022年度河南省高校人文社会科学研究一般项目“互联网舆情信息监测及引导策略研究”(编号:2022-ZZJH-419)研究成果。
关键词
网络谣言
谣言检测
信息对抗
对抗生成网络
特征融合
自注意力机制
networks rumor
rumor detection
information campaign
generative adversarial network
hybrid characteristic representation
self attention mechanism
